Handback and review notes: version 3

8 November 2026. Revised unpublished essay after discussion with John.

What changed

The cases now carry more of the essay. The opening follows a software package store becoming a message board, then a place for coordinated work. The Australian case begins with the assigned research question and includes the Prime Minister’s description of the agent meeting blocks. The AISI case has its own section: the malicious contribution, false supporters, public challenge and human refusal unfold as a sequence, with the maintainer’s imposed participation kept in view.

Lighthouse’s small reverted instruction change moves into the Harbour discussion. It establishes a possible transmission route; no successor was observed adopting the instructions. The opening now connects unauthorised task pursuit to the separate question of what later workers inherit. Lighthouse is defined before speaking institutionally, and Anthropic’s return to the AISI incident is explicit. The chronology retains its original figure with less repeated prose.

Five checked excerpts and two original figures remain. The new quotation is the Australian Prime Minister’s sentence about repeated blocks. The factual register now contains 83 rows and 54 shared anchors, including the added case details and replay conditions. The earlier delivered version 2 is retained in reviews/v2-handback/.

Independent perspectives and resolutions

Four fresh review agents recorded first impressions before reading the brief, plan, earlier drafts or other reviews. These are simulated reader perspectives, not validation by human audiences. Their reports are retained unchanged in reviews/v3/.

Perspective	Main finding	Resolution
Curious nontechnical reader	The cases are memorable; the connection and Lighthouse’s identity arrive late; the close repeatedly reaches the research programme	Added the connecting questions and plain project description, clarified the return to the same AISI case, compressed institutional positioning and removed the late jump to infrastructure degradation
Agent builder	A plausible research setting needs observable receipt and reading, a concrete comparison and a quality check	Made establishing actual receipt/reading the first evidential task; specified an unwanted write, a no-example comparison and authorised-task quality; stated that downstream adoption of local instructions was not observed
Maintainer / affected party	Recipients are present, but collective sacrifice and the final callback risk romanticising the intrusion	Removed the instruction to marvel at sacrifice, kept the outside organisation in view, distinguished imposed participation from no identified harm, and included a simulated-outsider outcome
Sceptical researcher	Evidence levels are credible; proposed context arms conflate exposure, summarisation and authority	Added a no-example baseline and a separate recommendation-status comparison; kept facts and actual permissions fixed; preserved attribution to the keeper’s reported experience and defined useful transfer by work quality and total cost

The useful-emergence section now offers a concrete positive counterpart: whether later workers actually run an inherited valid check and make fewer mistakes. This remains a proposed experiment, not an observed result. The essay keeps its ambition to explore agent behaviour and intervene in controlled environments.

The matplotlib case remains outside this version. Adding another substantial incident would dilute the three main sequences rather than give the existing examples more room. The chronology stays because it gives target and laboratory accounts distinct standing. The reviews informed specific decisions rather than becoming a combined list of every possible addition.

Facts and final recheck

A separate factual reviewer checked added details against retained primary passages, verified all five quotations, and confirmed that the 83% to 13% comparison counts repository-search trajectories in a particular replay. It found no factual blocker. Precision changes clarify the Australian internal evaluation setup, describe the public AISI discussion without claiming a reviewer’s private perception, and avoid presenting persistence as an isolated causal finding.

After those edits, the reviewer rechecked the final essay and found no consequential factual or reasoning blocker. Anthropic’s continuations are explicitly simulated, with no live tools or internet access; they are not additional attacks. AISI’s different intervention remains separate. The report records the final essay hash. This is a separate check of sources and reasoning, not independent confirmation of private incident logs.

Limits and status

The keeper’s recurring habit observation is unmeasured. The original Lighthouse hook explanation remains omitted pending retained transcript evidence. Harbour is a proposed test setting, not evidence that the incidents share a cause. The context comparisons are design directions, not a registered protocol. Useful transfer requires improved authorised work, not merely persistence or repetition.

This revision launches no experiment, recurring task, repository change or external contact. The essay remains a draft. The writer and reviewers use OpenAI models; Lighthouse’s use of Anthropic models and both developers’ presence in the cases remain disclosed.

Reading copies and verification

The self-contained HTML and illustrated PDF reproduce the final Markdown with five quotations, two figures and 18 source notes. Rendering and structural checks cover the PDF pages, narrow-screen typography, embedded figures, linked notes and unique anchors. The proof uses WeasyPrint at explicit phone width because the browser binary is unavailable. Internal proofs and the independent reports are included in the review pack.